def inference(opt):
device = check_envirionment(opt.use_cuda)
model = load_test_model(opt, device)
audio_paths, script_paths = load_data_list(opt.data_list_path, opt.dataset_path)
target_dict = load_targets(script_paths)
testset = SpectrogramDataset(audio_paths=audio_paths, script_paths=script_paths, sos_id=SOS_token, eos_id=EOS_token,
target_dict=target_dict, opt=opt, spec_augment=False, noise_augment=False)
evaluator = Evaluator(testset, opt.batch_size, device, opt.num_workers, opt.print_every, opt.decode, opt.k)
evaluator.evaluate(model)
def split_dataset(opt, audio_paths, script_paths):
"""
split into training set and validation set.
Args:
opt (ArgumentParser): set of options
audio_paths (list): set of audio path
script_paths (list): set of script path
Returns: train_batch_num, train_dataset_list, valid_dataset
- **train_time_step** (int): number of time step for training
- **trainset_list** (list): list of training dataset
- **validset** (data_loader.MelSpectrogramDataset): validation dataset
"""
target_dict = load_targets(script_paths)
logger.info("split dataset start !!")
trainset_list = list()
train_num = math.ceil(len(audio_paths) * (1 - opt.valid_ratio))
total_time_step = math.ceil(len(audio_paths) / opt.batch_size)
valid_time_step = math.ceil(total_time_step * opt.valid_ratio)
train_time_step = total_time_step - valid_time_step
base_time_step = train_time_step
if opt.spec_augment:
train_time_step += base_time_step
if opt.noise_augment:
train_time_step += base_time_step
train_num_per_worker = math.ceil(train_num / opt.num_workers)
# audio_paths & script_paths shuffled in the same order
# for seperating train & validation
tmp = list(zip(audio_paths, script_paths))
random.shuffle(tmp)
audio_paths, script_paths = zip(*tmp)
# seperating the train dataset by the number of workers
for idx in range(opt.num_workers):
train_begin_idx = train_num_per_worker * idx
train_end_idx = min(train_num_per_worker * (idx + 1), train_num)
trainset_list.append(
SpectrogramDataset(audio_paths[train_begin_idx:train_end_idx],
script_paths[train_begin_idx:train_end_idx],
SOS_token,
EOS_token,
target_dict=target_dict,
opt=opt,
spec_augment=opt.spec_augment,
noise_augment=opt.noise_augment,
dataset_path=opt.dataset_path,
noiseset_size=opt.noiseset_size,
noise_level=opt.noise_level))
validset = SpectrogramDataset(audio_paths=audio_paths[train_num:],
script_paths=script_paths[train_num:],
sos_id=SOS_token,
eos_id=EOS_token,
target_dict=target_dict,
opt=opt,
spec_augment=False,
noise_augment=False)
logger.info("split dataset complete !!")
return train_time_step, trainset_list, validset